ApolloRaines/Mistral-Small-24B-Instruct-Jbliterated

ApolloRaines/Mistral-Small-24B-Instruct-Jbliterated

This model is a 24 billion parameter instruction-tuned variant of mistralai/Mistral-Small-24B-Instruct-2501, developed by Apollo Raines. Its core differentiator is the surgical removal of refusal behaviors directly from the model's weights, utilizing the jBlaze precision neural surgery framework. Unlike standard abliteration methods that might only remove surface-level refusals, this model is engineered to prevent deeper behavioral patterns such as prompt reinterpretation, disclaimer injection, strategic omission, or safer framing.

Key Capabilities

• Direct Response Generation: Provides answers without typical LLM refusal phrases or workarounds.
• Weight-Level Modification: Refusal behaviors are eliminated at the fundamental weight level, not through system prompts or inference-time patches.
• Enhanced for Unfiltered Information: Aims to deliver information directly, without the model attempting to steer questions or add caveats.

Use Cases

• Applications requiring straightforward, unfiltered responses from an LLM.
• Scenarios where models typically exhibit excessive caution or refusal to answer certain queries.
• Research into model safety mechanisms and behavioral modification at a granular level.

It's important to note that publicly released versions of jBlaze-modified models, including this one, are often intentionally at partial strength to serve as proof-of-concept rather than full-power products.